English
By Topnet Engineering Team | Updated July 2026
✓ In Stock | Free Sample Available - see our sample policy | Volume & Tender Pricing - sales@topnetsystem.com
AI training clusters and hyperscale data centers are moving to 800G interconnects faster than any previous speed transition. GPU-to-GPU traffic in large language model training, east-west flows in spine-leaf fabrics, and NDR-class InfiniBand deployments all demand 800G optical transceivers today. This guide explains how to choose the right 800G module for your network: form factor (OSFP vs QSFP112), optical reach (SR8, 2xDR4, 2xFR4, 2xLR4), and thermal options (finned vs flat top).
An 800G optical transceiver converts eight electrical lanes of 100G PAM4 signaling into optical signals, delivering 800 Gbit/s aggregate bandwidth in a single pluggable module. Most 800G modules operate as 2x400G breakout-capable devices (for example 2xDR4 or 2xFR4 architectures with dual MPO or dual duplex-LC optical ports), which makes them ideal for connecting one 800G switch port to two 400G ports during migration. Key building blocks include 100G-per-lane SerDes, PAM4 modulation, and DSP-based signal recovery.
The two mainstream 800G form factors are OSFP (Octal Small Form-factor Pluggable) and QSFP112-based double-density packages. Both carry 8x100G electrical lanes; the differences are mechanical and thermal.
| Feature | 800G OSFP | 800G QSFP112 / QSFP-DD |
| Size | Larger housing | More compact, higher port density |
| Thermal capacity | Superior - supports integrated (finned) heatsink, up to ~18W | Relies on riding heatsink on the cage, typically up to ~15W |
| Typical ecosystem | NVIDIA InfiniBand NDR / Spectrum Ethernet AI clusters | Ethernet data center switches, backward-compatible cages |
| Backward compatibility | OSFP-RHS variants for some platforms | Accepts QSFP28/QSFP56 with adapters in many designs |
| Best for | High-power AI/GPU clusters | Space-constrained, high-density Ethernet fabrics |
Rule of thumb: if your switches and NICs are NVIDIA AI-cluster class, you will most likely need OSFP; for mainstream Ethernet data center switching, QSFP112 keeps density high and cages compatible. Topnet System supplies both: see our 800G OSFP & QSFP112 transceiver range.
| Type | Fiber | Wavelength | Reach | Typical use |
| 800G SR8 | MMF (OM3/OM4), MPO | 850nm VCSEL | up to 100m | In-row GPU cluster links, lowest cost per port |
| 800G 2xDR4 | SMF, dual MPO-12 APC | 1310nm | 500m | Leaf-spine within building, 2x400G-DR4 breakout |
| 800G 2xFR4 | SMF, dual duplex LC | CWDM4 1310nm band | 2km | Cross-building campus and larger halls |
| 800G 2xLR4 | SMF, dual duplex LC | CWDM4 | 10km | Data center interconnect (DCI), metro edge |
Start from your actual link-length distribution: the majority of AI-cluster links are under 100m (SR8 territory), leaf-spine runs typically fall inside 500m (2xDR4), while campus and DCI links justify 2xFR4 or 2xLR4. Mixing reach types per tier, rather than standardizing on one, usually cuts optics budget by 20-30%.
OSFP modules come in two mechanical tops: finned (integrated heatsink) for air-cooled switch cages designed for it (most NVIDIA switch ports), and flat top for cages with riding heatsinks or liquid-cooled systems, and for NIC ports such as ConnectX adapters. Ordering the wrong top is one of the most common 800G procurement mistakes - always confirm the cage type of both link ends. Topnet System offers finned and flat-top variants across the 800G family, including Sipho (silicon photonics) based versions for lower power.
1) Confirm form factor per device - switch side may be OSFP while NIC side is QSFP112. 2) Match reach with margin - include patch panel and connector losses in your budget. 3) Check fiber plant - MPO-12 APC polarity for DR4-class, duplex LC for FR4/LR4. 4) Verify power and cooling - 13-18W per port adds up across 64-port switches. 5) Test before volume - request samples and validate in your own chassis with your own traffic profile.
Yes - 2xDR4 and 2xFR4 modules split into two independent 400G links, so one 800G switch port can serve two 400G devices during migration.
Both use eight 100G optical lanes at 1310nm over SMF. DR8 presents a single MPO-16 interface as one 800G link; 2xDR4 presents dual MPO-12 ports that can operate as two independent 400G-DR4 links.
Finned for most air-cooled switch cages (e.g. NVIDIA Quantum/Spectrum switch ports); flat top for riding-heatsink cages, liquid-cooled systems, and NIC ports. Confirm per device before ordering.
1.6T OSFP224 (200G per lane) is emerging for XDR-class fabrics, but 800G remains the volume workhorse through the next several hardware generations, and 800G optics will keep serving the 400G/800G tiers of 1.6T networks. Deploy 800G now; plan cabling (MPO-16/APC choices) with 1.6T in mind.
Topnet System provides samples under a refundable-deposit, 30-day testing program - see the Free Sample & Testing Policy. Full test reports (eye diagram, TX/RX power, temperature cycling) are available on request.
Topnet System (ISO certified since 2008) manufactures the full 800G family - OSFP SR8, 2xDR4/2xDR4+/2xDR4++, 2xFR4, 2xLR4, Sipho variants, and QSFP112 2xSR4 - engineered for AI data centers and GPU compute clusters. Browse the 800G OSFP QSFP112 category, or email sales@topnetsystem.com for datasheets, samples and volume pricing.